An Optimized Approach for Real-Time License Plate Detection and Tracking Using YOLO-NAS and SORT With Comparative Analysis of YOLO Variants

Abstract

This paper introduces an innovative deep learning model to enhance license plate recognition systems, using YOLO-NAS for accurate detection and recognition in real-world settings. The model, trained on diverse datasets, incorporates a novel quantization-friendly basic block, addressing previous YOLO model constraints. Performance is boosted with advanced training and post-training quantization techniques. Concurrently, YOLOv8 classifies vehicle types, and the SORT algorithm assigns unique IDs to vehicles, ensuring seamless license plate association, with data stored in a CSV file. EasyOCR recognizes and displays alphanumeric characters on plates. YOLO-NAS achieves 90.2% accuracy, combining high speed and precision, significantly advancing automated license plate recognition and enhancing traffic management and security applications.